Compound-Protein Interaction Prediction with Sparse Perturbation-Aware Attention

Compound-Protein Interaction (CPI) prediction is a crucial task in drug discovery. Modern CPI prediction models are mostly based on the attention mechanism. However, the attention scores are often inaccurate, i.e., functionally irrelevant substructures can still receive moderate attention scores, and attention scores can not distinguish compounds with similar structural topology but different pharmacological properties. We propose SPACPI to address this problem from three perspectives, i.e., (1) identifies important compound substructures by integrating auxiliary information from molecular fingerprints, (2) determines important compound atoms by learning each atom’s tolerance to different perturbation amplitudes, (3) obtains more robust model parameters by focusing on the topK important atoms. Experiments on two benchmark datasets and two label-reversal datasets show that SPACPI outperforms the state-of-the-art CPI prediction model with an average increase of 5.02% across different datasets and evaluation metrics. Visualization verifies that SPACPI can produce more accurate and explainable predictions.